The collection interface is something applications and toolkits get
"for free" from AT-SPI2. However, applications/toolkits that provide
their own implementation of AT-SPI2 (e.g. WebKitGtk) presumably need
to also implement AtspiCollection themselves, or do something so that
the default AT-SPI2 implementation is used. Currently, in WebKitGtk
version 2.36, a non-implemented error occurs when querying the
collection interface on the document. We were not handling this error,
causing Orca to be silent when structural navigation commands were
used. This commit handles the error. It cannot solve the missing
implementation, however. Thus in WebKitGtk apps,  Orca will now announce
things like "no more headings" when there are in fact headings present.

See issue #244

When generating ancestors, we typically stop at large containers so as
not to be too chatty. We were treating DOCUMENT_FRAME as such a container,
primarily because it was how web documents were exposed to us (prior to
the creation of DOCUMENT_WEB). Now that we have DOCUMENT_WEB, we should
be able to distinguish ARIA's document role (DOCUMENT_FRAME), from a web
document. And because we cannot control where an author might choose to
use this role, we should not take it as a reason to stop generating
ancestors. Example: The FluentUI react dialog applies the document
role to a div which stands between the buttons and the dialog.

Fixes issue 250

* Announce entering and leaving list
* Announce item count (based on the number of description list terms)
* Append the role name at the end of each description list term in
  order to make it possible to distinguish terms from their descriptions
  in confusing lists, while still making it possible for users to cut
  off that speech in cases where the distinction is obvious.
* Similarly append the role name at the end of each description list
  value.

See issue #248

When a newly-expanded/shown listbox appears and focus is given to
a child option, we generate ancestors. That presentation includes
the name of the ancestor listbox.

In some ARIA authoring, the name of the listbox is the same as the
previously-focused object (e.g. a button or collapsed combobox).
Repeating this name is unnecessarily chatty. This commit checks for
that condition and returns early for the listbox name generation.

Possibly fixes issue #252.

If a user is tabbing among links inside an ARIA grid, we were not
speaking the headers and positions of column headers. In order to fix
this completely, we need to do two things: 1) Generate ancestors and
2) Adjust what information gets generated for the cell information.
The second might have undesired side effects, so the changes are being
split. This commit only addresses the first.

See issue #253

The 'focused' formatType has been serving dual purposes, specifying
what to present when the object is already focused and something
changed and what to present about an ancestor of a newly focused
object. This approach was likely done because the information in
both cases is expected to be minimal, and the set of roles for
which we might want distinct already-focused info and as-ancestor
info small to non-existent.

It turns out that small may be true; non-existent isn't. Consider focus
changing among items (e.g. links) descending from ARIA gridcells. If the
column changes, we want to include some of the new cell/column info. And
that info will be different from what we present when some state
changes in the currently-focused Gtk table cell.

This commit merely adds the capability of an ancestor formatType
because having it makes sense in general. It falls back on the focused
formatType, so no end-user-observable behavior should result from this
commit. A subsequent commit will use this new formatType to solve the
aforementioned table cell problem. And as other such instances become
noticed, we can start decoupling "focused" from "ancestor."

See issue #253.

This commit causes us to present the row and/or column header when
the user moves focus to a descendant of a cell in a different row
and/or column.

See issue #253.

The logic to determine whether or not the row and/or column header
has changed relies on the values stored when the location changes.
We were only storing those values when the new location was a cell
(or cell-like thing). If instead the new location is a focusable
descendant of the cell, we are at risk of duplicating header info.

This commit fixes that by looking for the most immediate ancestor
cell and storing its coordinates if found.

See issue #253.

This commit causes us to present the row and/or column index when
the user moves focus to a descendant of a cell in a different row
and/or column.

See issue #253.

Because the time element has various punctuation marks in it, user
agents sometimes expose it as being divided into multiple sentences.
Given a time element inside of a link while doing a SayAll by
sentence, we can theoretically repeat "link" after each substring
reported to us as a full sentence. Alternatively we might say "link"
only once, but somewhere in the middle rather than at the end.

This commit solves that problem by always treating the time element
as a single sentence, rather than asking user agents for its sentences.

Fixes issue #247